How niravrvaghasiya/skill-intelligence-engine works
Skill Intelligence Engine (SIE) is a corpus-agnostic library, CLI, and small API for finding and explaining relevant skills in SKILL.md collections. Its main workflow loads a corpus, chunks skill text, retrieves candidates with BM25 and optionally dense search, fuses and optionally reranks them, then reports confidence and a route/clarify/abstain decision. Separate capabilities compose multi-intent plans and derive learning paths from declared skill relationships. The graph follows documented entry points and sampled ownership; some subsystem wiring is based on README capabilities and the file tree.
Main entry points: router.py CLI and router · engine.py Engine · api.py HTTP API · dense.py Dense index
Components 18
Interfaces and orchestration
- HTTP API
sie/api.py - Engine
sie/engine.py - CLI and router
sie/router.py
Retrieval and ranking
- Text chunking
sie/chunking.py - BM25 index
sie/index/sparse.py - Dense index
sie/index/dense.py - Persisted vectors
- Rank fusion
sie/index/fuse.py - Cross-encoder reranker
sie/rerank.py
Corpus and skill graph
- Corpus ingestion
sie/ingest.py - Loaded skills
sie/corpus.py - Relationship graph
sie/graph/build.py
Planning and results
- Confidence decision
sie/confidence.py - Multi-intent composition
sie/compose.py - Learning paths
sie/graph/paths.py - Ranked skill results
sie/models.py
Other components
- Task requester
- SKILL.md corpus
Connections 19
- Task requester → HTTP API: submits task
- HTTP API → Engine: dispatches
- Task requester → CLI and router: runs query
- SKILL.md corpus → Corpus ingestion: loads files
- Corpus ingestion → Loaded skills: creates skills
- Engine → CLI and router: uses retrieval
- Engine → Relationship graph: builds graph
- CLI and router → Text chunking: chunks skills
- CLI and router → BM25 index: searches BM25
- CLI and router → Dense index: searches vectors
- Dense index → Persisted vectors: reads and writes
- CLI and router → Rank fusion: fuses rankings
- CLI and router → Cross-encoder reranker: reranks candidates
- CLI and router → Confidence decision: assesses route
- CLI and router → Ranked skill results: returns ranking
- Engine → Multi-intent composition: plans intents
- Engine → Learning paths: requests path
- Relationship graph → Learning paths: supplies relations
- HTTP API → Ranked skill results: returns response